Leaderboard Submission: CarMRI (PointNet2D + BiLSTM)

[Adarsh-Roy-DF]

Award

Miltos Petridis Memorial Trophy — Best Student Application Paper at AI-2025, the 45th SGAI International Conference on Artificial Intelligence, BCS, Cambridge University (December 2025)

Summary

• Submission of CarMRI, a lightweight slice-based surrogate model for aerodynamic drag coefficient (Cd) prediction on DrivAerNet++
• Architecture: PointNet2D + BiLSTM — 3D point clouds are sliced into 80 sequential 2D cross-sections (inspired by MRI), encoded by PointNet2D, and sequenced through a Bi-LSTM
• Evaluated on the official DrivAerNet++ test split (1,158 samples)

Test Results

Metric	Value
MSE	6.60 × 10⁻⁵
MAE	6.111 × 10⁻³
Max MAE	4.50 × 10⁻²
R²	0.9525
Inference Latency	0.025 s/sample
Parameters	2.80M
Total Inference Time	28.95s for 1158 samples

Links

• Paper (Springer Nature LNCS): https://doi.org/10.1007/978-3-032-11442-6_5
• Paper (arXiv): https://arxiv.org/abs/2601.02112
• Code & Weights: https://github.com/Adarsh-Roy/cd_prediction